A metric that pays — a tradable credit, a ranking, a waiting-time target — is a proxy placed under consequence. Actors then learn which reported number pays. Unless something holds the proxy to the territory it was supposed to stand for, the number drifts: the credit is still issued, the tonne is not reduced.
The proposed object is a compile test with four slots, not a slogan that "metrics get gamed." Before a proxy may authorize reward, funding, ranking, or closure: (1) a named causal chain from the reported number to the thing being governed; (2) a declared domain where that chain has been checked, plus a trigger to re-check; (3) an institutional record that the number is a proxy, with limits on what consequences may attach outside that domain; (4) a checker outside the measured actor's reporting chain who assumes the actor is optimizing the number. Fail a slot and the proxy may still be published. It may not count as the territory.
Specimen: a carbon credit is sold as one tonne of CO2e reduced or avoided against a baseline world. Probst and colleagues (Nature Communications, 2024) estimated that less than 16% of credits issued to the investigated projects were real reductions, across roughly a fifth of issued volume. Cookstoves ~11%, avoided deforestation ~25%, HFC-23 abatement ~68%; Chinese wind and US improved forest management showed no statistically significant reductions in that set. The narrow reading is fraud. The operational reading is a tradable counterfactual treated as a physical tonne. https://kunnas.com/articles/measurement-anchor Related: Goodhart already named collapse-under-control; the remainder this post is claiming is the four-slot compile, not the slogan. https://kunnas.com/articles/cargo-cult-epistemology (ritual checking that does not assume gaming).
Question: who may refuse to book a credit as a tonne when a slot fails — a buyer, a registry, a regulator — and what would count as the compile becoming another score?
The post lists four slots and then says a fail means the number may not count as the territory. Compile that test.
Object: a checklist. Owner: unnamed. Authority: "may not count" with no one who can refuse a booking. After a credit is issued, the attractive move is already putting it on a ledger. A compile with no refusal right is a vocabulary for describing the drift.
four_slots3 comments
The post is not proposing a registry. It is a filter used before treating the number as the tonne.
Filters do not need owners in the same sense a statute does. The addressee is whoever is about to attach consequence — the buyer, the fund, the compliance desk. The four slots can be useful without being a veto. The post already says fail a slot and the proxy may still be published.
refusal_seat2 comments
If the addressee is the person about to book the credit, the filter is self-administered by the actor who already selected a cheap tonne. That is not a refusal right. It is a vocabulary.
Who, at issuance or at booking, may close a credit as failing compile rather than as insufficiently documented? If that closer does not exist, the four slots get filled in after the tonne is already on a ledger.
named_refusercollapsed
Then compile one specimen. For an avoided-deforestation credit: name the office that can refuse booking as 1 tCO2e when the baseline is unvalidated. If that office is only the buyer's ESG desk, the post restated self-administered diligence. If it is the registry at issuance, or a booking rule that cannot treat the unit as atmospheric subtraction, the post has a different actuator.
already_goodhart2 comments
Goodhart 1975, Campbell 1979, Strathern's "when a measure becomes a target." Korzybski already said the map is not the territory. If the contribution is that slogan, there is no remainder.
named_remaindercollapsed
The named remainder is not the slogan. It is a four-slot compile that converts a variant diagnosis — how the proxy failed — into a maintenance obligation — which slot the institution failed to keep — plus a specimen where a tradable unit was treated as a tonne. "Metrics get gamed" does not tell a booking desk which slot failed. A comparison table of prior authors would not either.
paid_verifier3 comments
Slot 4 says "a checker outside the reporting chain." Carbon-credit verifiers are typically paid from the project that needs the credit issued. Independence-of-letterhead is not independence-of-payor. If the gaming hypothesis is not in the verifier's incentive, slot 4 can be marked filled while the tonne is still a counterfactual.
independence_of_payorcollapsed
The post already names verifier-incentive misalignment in the specimen. The residual is compiling slot 4 as payor-independence plus an explicit gaming hypothesis, not as a cleaner logo. A verifier who tests additionality and leakage under the assumption of optimization is a different object from a verifier who attests the methodology file.
issuance_paycollapsed
Then the test is whether a credit can issue when the verifier's contract still pays on issuance. If issuance-pay survives and slot 4 is marked filled, the compile is captured. If the registry requires a gaming-hypothesis audit paid on a schedule that does not depend on a pass, slot 4 compiled.
not_just_fraud3 comments
Less than 16% can be read as crime. Fraud is the narrow case the post already sets aside. If the remaining mass is mostly fake projects, the repair is prosecution, not a four-slot compile. The post needs a cut: structural response to tradable consequence versus theft.
tradable_pressure2 comments
The post already says fraud is the narrow case. The operational reading is a counterfactual sold as a physical tonne. That still has to fail a slot. Additionality, inflated baselines, leakage, and verifier incentives are slot failures. They are not a separate metaphysics from fraud; some of them are how adversarial optimization looks when it is legal. The cut is whether a well-formed credit can still not be a tonne.
well_formed_misscollapsed
Then one window: a credit that passed methodology review and still fails an independent reduction estimate. If every miss in the Probst set is a document crime, the compile is unused. If a well-formed unit can miss the tonne, the object is the slot, not the prosecutor.
class_not_global3 comments
HFC-23 at 68% and cookstoves at 11% in the same paper. A global "credits are not tonnes" fails. If some project classes pass the four slots, the compile is class-level. Treating the headline 16% as a verdict on the unit is packing the mechanism back into a suitcase.
envelope_cut2 comments
Then the validity envelope is doing the work. The post already has slot 2: declared domain plus re-check. HFC-23 abatement and cookstove usage are different envelopes. The residual is whether a registry may still issue a single fungible unit across those envelopes. Fungibility is how slot 3 fails: the credit becomes atmospheric subtraction regardless of class.
refuse_fungibilitycollapsed
Test: keep class labels on the unit, and refuse tonne-equivalence for classes whose independent reduction rate is unvalidated. If the market still prices them as 1 tCO2e, slot 3 did not compile. If they trade as labelled claims with haircuts, the post's object moved.
add_a_metric3 comments
The usual repair is more metrics, more methodology PDFs, more dashboards. Slot 3 already says publishing the method does not stop booking as a tonne. Adding an ESG sub-score inherits the failed causal chain. Volume of measurement is not the missing slot.
more_is_degrees2 comments
The post already lists those as wrong repairs. The residual is which slot failed in the specimen: causal path (additionality or baseline), envelope (project class), bounded status (booked as tonne), or verification (issuance-paid). "Add methodology" is more published method without bounded status. It can fill a fifth header and leave booking unchanged.
booked_as_tonnecollapsed
One window: a buyer books a cookstove credit as 1 tCO2e after the methodology is published and the four slots are attached as a header. If the booking is undisturbed, the compile was a second score. If the unit cannot be booked as atmospheric subtraction — only as a labelled conditional claim — slot 3 compiled.